AI Systems Architecture

Quality Gate System

Every pipeline output passes through weighted scoring rubrics. Agents don't just generate — they evaluate, score, and enforce thresholds before anything ships.

3x weight

Spec Compliance

Does the code implement what was specified?

Every endpoint, page, and feature from spec exists Auth gates match spec requirements Environment variables documented Gap analysis: spec-to-code AND code-to-spec

2x weight

Copy Integration

Is marketing copy used verbatim — not paraphrased?

All Hype agent copy appears word-for-word in UI No placeholder text ("Lorem ipsum", "Coming soon") CTAs match brand voice document

2x weight

Code Quality

Logic correctness, error handling, maintainability.

No logic errors or unreachable code paths Error handling present at system boundaries No hardcoded values that should be config

1x weight

Security Basics

OWASP Top 10 coverage with attacker persona.

No hardcoded secrets or API keys Input validation, parameterized SQL XSS prevention, CSRF tokens where needed Mandatory: "If I wanted free usage, what would I try?"

Simulated Score Breakdown

Spec Compliance

Copy Integration

Code Quality

Security Basics

Weighted Score 86

PASS — threshold is 80%. Scores 60-79% trigger revision loop. Below 60% = critical.

Why Async/await trace is mandatory. For every async call: is it awaited? What happens if the Promise resolves as undefined? Does the conditional treat undefined as truthy? The review traces actual execution — not intended behavior.

3x weight

AI Slop Detection

Automated grep scan before human-level review. Any match = instant zero.

Regex scan runs BEFORE reading for meaning Each hit quoted with exact line reference Specific human replacement provided per hit

2x weight

Brand Voice Alignment

Tone matches voice doc. Words-to-use present, words-to-avoid absent.

Voice adjectives reflected in sentence structure Words-to-use list checked for presence Words-to-avoid list checked for absence

2x weight

Conversion Effectiveness

Does the copy actually sell?

Clear value prop in first 10 words Compelling headline with specific benefit CTA is action-specific, not generic Social proof present and credible

1x weight

SEO & Technical

Meta descriptions, title tags, heading hierarchy.

Meta description under 160 characters Title tag under 60 characters Primary keyword in H1 Heading hierarchy: H1 → H2 → H3, no skips

Auto-Fail Phrases (Any Present = Score 0)

Grep-based scan runs before semantic review. These phrases indicate AI-generated copy that wasn't edited by a human.

game-changer revolutionize unlock the power in today's fast-paced world seamlessly effortlessly cutting-edge state-of-the-art leverage empower supercharge take to the next level say goodbye to it's not just X it's Y in a world where comprehensive solution paradigm synergy best-in-class

Why AI-generated copy that reads like AI-generated copy destroys trust. The auto-fail list catches the most common tells. If a marketing agent produces slop, it gets scored zero and re-dispatched — no exceptions.

3x weight

Cross-Agent Coherence

Does the product match the spec? Does the landing page describe the real product?

Product behavior matches Blueprint spec Landing page describes actual features, not aspirational ones FAQ answers match real product behavior Help docs reference actual UI elements

2x weight

Completeness

Nothing missing from the full delivery.

All endpoints implemented and responding All pages from spec exist with real content Forms have validation and error states Legal pages in footer, meta tags present, 404 page exists

2x weight

User Journey

Can a stranger understand and use this in 5 seconds?

Value prop clear within 5 seconds of page load Primary CTA above the fold Happy path works end-to-end without errors Error states are graceful, not stack traces

1x weight

Launch Readiness

Would you show this to 100 strangers?

No obvious "AI-built" tells Performance acceptable (no 5s+ load times) Mobile layout doesn't break Brand colors, typography, spacing consistent

Gate Reflect-Project is the final launch gate. If it fails, the pipeline doesn't ship. Period. This catches the coherence gaps that individual agent reviews miss — the landing page promising a feature that Forge didn't build, or help docs referencing a UI that changed.

Prompt Architecture

Three-layer progressive disclosure keeps context windows efficient. Skills load only what's needed, when it's needed.

L1 Metadata — Always in Context ~100 tokens ▶

Name and description only. Loaded into every conversation. The description determines when the skill triggers — it's optimized for recall, not readability.

name: reflection--code-review
description: Deduction-based code quality scoring with weighted
  rubrics, async/await trace analysis, and attacker
  persona security review. Triggers on: code review,
  quality check, "is this code safe", or any Forge
  output before pipeline continues.

Design Description is a trigger classifier, not documentation. It's tuned so the model invokes the skill at the right moment — like training a retrieval function.

L2 Skill Body — Loaded on Invocation <500 lines ▶

Full instructions injected when the skill is invoked. Structured as methodology, not rigid rules. Explains why behind every instruction so the model can handle edge cases.

# From the actual Reflect-Code skill body:

## Async/Await Trace (Critical)
For every async function call:
1. Is the call `await`ed?
2. Is the return value used?
3. If missing `await`, trace exact runtime behavior:
   - What does Promise object look like used synchronously?
   - What happens when you access `.allowed` on pending Promise?
   - `!undefined` is `true` — does the conditional pass?
   - Write ACTUAL execution path, not intended path

## Attacker Persona (Mandatory Output)
After line-by-line review, answer:
1. "If I wanted unlimited free usage, what would I try?"
2. "If I wanted to use without paying?"
3. "Which identifiers can be rotated, spoofed, omitted?"
4. "Which safety checks exist but aren't called/awaited?"

Why Theory of mind over rigid MUSTs. LLMs are smart — if you explain why a check matters, the model generalizes to edge cases instead of pattern-matching to specific examples.

L3 Bundled Resources — Loaded as Needed Unlimited size ▶

Scripts, templates, reference docs. Only loaded when the skill explicitly reads them during execution. Keeps context window clean.

# Skill directory structure
skill-name/
├── SKILL.md           # L2: instructions (required)
│   ├── YAML frontmatter  # L1: name + description
│   └── Markdown body     # L2: methodology
└── Bundled Resources  # L3: loaded on demand
    ├── scripts/          # Executable code
    ├── references/       # Docs read during execution
    └── assets/           # Templates, fonts, icons

Spec Handoff: 16-Section Structured Output

Every product spec follows the same structure. The Blueprint agent fills all 16 sections, then hands off to Forge with: "Read SPEC.docx and CLAUDE.md. Begin Phase 1."

Cover + Summary Table Architecture + Stack Database Schema API Endpoints Core Business Logic Frontend Key Pages Environment Variables Build Order (5-7 phases) Pricing & Revenue Key Decisions Credentials Setup File/Folder Structure OAuth Callback URLs Security Headers Deployment Config Error Response Envelope

Design Build order is phased so each phase is independently testable. Phase 1 is always: project init + models + database + auth. Last phase is always: deploy + README. This means if any phase fails, you know exactly where.